initial commit, should there be a ResponseReasoningPartAdded type?

Signed-off-by: Andrew Xia <axia@meta.com>

Author: qandrew

vllm/entrypoints/openai/protocol.py

@@ -31,6 +31,8 @@
     ResponseReasoningTextDeltaEvent, ResponseReasoningTextDoneEvent,
     ResponseStatus, ResponseWebSearchCallCompletedEvent,
     ResponseWebSearchCallInProgressEvent, ResponseWebSearchCallSearchingEvent)
+from openai.types.responses.response_reasoning_item import (
+    Content as ResponseReasoningTextContent)
 
 # Backward compatibility for OpenAI client versions
 try:  # For older openai versions (< 1.100.0)
@@ -260,26 +262,6 @@ def get_logits_processors(processors: Optional[LogitsProcessors],
                                            ResponseReasoningItem,
                                            ResponseFunctionToolCall]
 
-StreamingResponsesResponse: TypeAlias = Union[
-    ResponseCreatedEvent,
-    ResponseInProgressEvent,
-    ResponseCompletedEvent,
-    ResponseOutputItemAddedEvent,
-    ResponseOutputItemDoneEvent,
-    ResponseContentPartAddedEvent,
-    ResponseContentPartDoneEvent,
-    ResponseReasoningTextDeltaEvent,
-    ResponseReasoningTextDoneEvent,
-    ResponseCodeInterpreterCallInProgressEvent,
-    ResponseCodeInterpreterCallCodeDeltaEvent,
-    ResponseWebSearchCallInProgressEvent,
-    ResponseWebSearchCallSearchingEvent,
-    ResponseWebSearchCallCompletedEvent,
-    ResponseCodeInterpreterCallCodeDoneEvent,
-    ResponseCodeInterpreterCallInterpretingEvent,
-    ResponseCodeInterpreterCallCompletedEvent,
-]
-
 
 class ResponsesRequest(OpenAIBaseModel):
     # Ordered by official OpenAI API documentation
@@ -1978,6 +1960,72 @@ def from_request(
         )
 
 
+# TODO: this code can be removed once
+# https://github.com/openai/openai-python/issues/2634 has been resolved
+class ResponseReasoningPartDoneEvent(OpenAIBaseModel):
+    content_index: int
+    """The index of the content part that is done."""
+
+    item_id: str
+    """The ID of the output item that the content part was added to."""
+
+    output_index: int
+    """The index of the output item that the content part was added to."""
+
+    part: ResponseReasoningTextContent
+    """The content part that is done."""
+
+    sequence_number: int
+    """The sequence number of this event."""
+
+    type: Literal["response.reasoning_part.done"]
+    """The type of the event. Always `response.reasoning_part.done`."""
+
+
+# TODO: this code can be removed once
+# https://github.com/openai/openai-python/issues/2634 has been resolved
+class ResponseReasoningPartAddedEvent(OpenAIBaseModel):
+    content_index: int
+    """The index of the content part that is done."""
+
+    item_id: str
+    """The ID of the output item that the content part was added to."""
+
+    output_index: int
+    """The index of the output item that the content part was added to."""
+
+    part: ResponseReasoningTextContent
+    """The content part that is done."""
+
+    sequence_number: int
+    """The sequence number of this event."""
+
+    type: Literal["response.reasoning_part.added"]
+    """The type of the event. Always `response.reasoning_part.added`."""
+
+
+StreamingResponsesResponse: TypeAlias = Union[
+    ResponseCreatedEvent,
+    ResponseInProgressEvent,
+    ResponseCompletedEvent,
+    ResponseOutputItemAddedEvent,
+    ResponseOutputItemDoneEvent,
+    ResponseContentPartAddedEvent,
+    ResponseContentPartDoneEvent,
+    ResponseReasoningTextDeltaEvent,
+    ResponseReasoningTextDoneEvent,
+    ResponseReasoningPartAddedEvent,
+    ResponseReasoningPartDoneEvent,
+    ResponseCodeInterpreterCallInProgressEvent,
+    ResponseCodeInterpreterCallCodeDeltaEvent,
+    ResponseWebSearchCallInProgressEvent,
+    ResponseWebSearchCallSearchingEvent,
+    ResponseWebSearchCallCompletedEvent,
+    ResponseCodeInterpreterCallCodeDoneEvent,
+    ResponseCodeInterpreterCallInterpretingEvent,
+    ResponseCodeInterpreterCallCompletedEvent,
+]
+
 BatchRequestInputBody = Union[ChatCompletionRequest, EmbeddingRequest,
                               ScoreRequest, RerankRequest]
 

vllm/entrypoints/openai/serving_responses.py

@@ -58,6 +58,8 @@
                                               InputTokensDetails,
                                               OutputTokensDetails,
                                               RequestResponseMetadata,
+                                              ResponseReasoningPartAddedEvent,
+                                              ResponseReasoningPartDoneEvent,
                                               ResponsesRequest,
                                               ResponsesResponse, ResponseUsage,
                                               StreamingResponsesResponse)
@@ -1280,14 +1282,13 @@ async def _process_harmony_streaming_events(
                         # Deal with tool call here
                         pass
                     elif previous_item.channel == "analysis":
+                        content = ResponseReasoningTextContent(
+                            text=previous_item.content[0].text,
+                            type="reasoning_text",
+                        )
                         reasoning_item = ResponseReasoningItem(
                             type="reasoning",
-                            content=[
-                                ResponseReasoningTextContent(
-                                    text=previous_item.content[0].text,
-                                    type="reasoning_text",
-                                ),
-                            ],
+                            content=[content],
                             status="completed",
                             id=current_item_id,
                             summary=[],
@@ -1301,6 +1302,15 @@ async def _process_harmony_streaming_events(
                                 content_index=current_content_index,
                                 text=previous_item.content[0].text,
                             ))
+                        yield _increment_sequence_number_and_return(
+                            ResponseReasoningPartDoneEvent(
+                                type="response.reasoning_part.done",
+                                sequence_number=-1,
+                                item_id=current_item_id,
+                                output_index=current_output_index,
+                                content_index=current_content_index,
+                                part=content,
+                            ))
                         yield _increment_sequence_number_and_return(
                             ResponseOutputItemDoneEvent(
                                 type="response.output_item.done",
@@ -1412,17 +1422,15 @@ async def _process_harmony_streaming_events(
                             ))
                         current_content_index += 1
                         yield _increment_sequence_number_and_return(
-                            ResponseContentPartAddedEvent(
-                                type="response.content_part.added",
+                            ResponseReasoningPartAddedEvent(
+                                type="response.reasoning_part.added",
                                 sequence_number=-1,
                                 output_index=current_output_index,
                                 item_id=current_item_id,
                                 content_index=current_content_index,
-                                part=ResponseOutputText(
-                                    type="output_text",
+                                part=ResponseReasoningTextContent(
                                     text="",
-                                    annotations=[],
-                                    logprobs=[],
+                                    type="reasoning_text",
                                 ),
                             ))
                     yield _increment_sequence_number_and_return(